Hi all,
I'm a student at Columbia University, and I'm taking a graduate course with
Alfred Aho, the author of the dragon book, on advanced compilers techniques.
I've been researching the pypy project in general, and rpython in particular,
and I'd like to ask you guys for some feedback on the current sketch of my
presentation. Aho has mentioned on several occasions that he is very excited to
receive my talk, and I'd like to get some feedback from you guys about it
before I put it forward to him.
So then, my talk will discuss rpython's approach to translation, and here is
the current outline:
- Compiling python to C is easy: just inline the implementation of every
opcode handler durr hurr hurr
- Ok, seriously, can you do it in a performant manner?
- Python has some semantics that make this difficult, in particular:
- opcodes are type-agnostic
- opcodes are high-level, they do high-level things with high-level
arguments. eg. the BUILD_CLASS opcode
- opcodes include namespace operations
- This type-agnostic bit is the real tricky part because C requires all
expressions to have a type, while python does not
- Vanilla cartesian product type inference doesn't really work because the
number of types is undecidable
- rpython gets around this by imposing a restriction on dynamic type creation.
- The details of the annotator are omitted due to time constraints
So then, the crux of my talk is the type annotator. I won't be going deeply
into the flow and object spaces in the talk, deferring to my paper. Is there
anything major I'm missing? Did I get anything horribly wrong?
Alex
_______________________________________________
pypy-dev mailing list
[email protected]
http://mail.python.org/mailman/listinfo/pypy-dev
